A Robotics Framework for Integrating Model-based Reasoning and Experience-based Learning
Abstract
In this project, we propose a framework that integrates model-based reasoning and experience- based learning for robots executing complex tasks in naval domains. To create a unified frame- work, the model-based and learning-based components need to be capable of communicating with each other. For example, the model-based reasoning needs to understand the effects on the model of executing a learned skill, and the learner needs to be able to parameterize its skills with elements from the internal world model and generalize the execution accordingly. This transfer of information is not trivial as learning and model-based reasoning methods employ fundamentally different representations. To reconcile the two approaches, we will develop representations that combine the core structure of the models with adaptive vector space embeddings from learning. Similar to the central nervous system in humans, there presentations will form a connection that grounds the robot???s internal world model in its external sensorimotor experiences. These shared representations will adapt and expand to incorporate novel concepts over time, while still maintaining the fundamental prior domain knowledge and interpretability of the original models. To create a seamlessly integrated framework, we will not only have the components communicate with each other, but also actively build on each other. In particular, the learning processes will extract and utilize prior domain knowledge from the models and planners, while the model- based reasoning processes will use the adaptability of the learning methods to expand the models and simplify the planning problems. By interweaving model-based reasoning and learning in this manner, the robot will continue to adapt to new experiences and benefit from long-term deployment. Similar to a human, the robot will be able to rely increasingly on prior knowledge from its own past experiences to learn new tasks and novel concepts more efficiently. The proposed frame work will be applicable to a wide variety of robot platforms and application areas. Our evaluation of the framework will focus on robots searching for objects of interest and clearing rooms. These tasks require the robots to learn a variety of skills for navigating unstructured environments as well as inspecting and manipulating objects.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 19, 2018
- Source ID
- N000141812775
Entities
People
- Maxim Likhachev
Organizations
- Massachusetts Institute of Technology
- Office of Naval Research
- United States Navy